In 2016 it was all about SaaS. Companies bought SaaS products quickly, relying on faith alone with no measurement plan. The same is happening with AI today. Companies bring in new AI tools for everything, AI assistants, code generation, without fully understanding how to measure performance outcomes and govern the cost.

What to do right now
- Track tokens consumed, API calls, and cost-per-developer-per-day from day one. That information becomes your baseline for understanding usage and leverage at renewal.
- Define your cost unit. Is it cost per PR, cost per accepted suggestion, or cost per resolved bug? Especially in PE-backed environments this matters. Investors understand unit economics, so speak their language.
- Evaluate each vendor rigorously. Usage patterns, models used, acceptance rate, latency, and model drift reports. Any vendor who won't share this data is a subscription, not a partner.
- Start with a limited number of engineers and measure. Roll out to others based on evidence, usage, and outcome.

Next steps
- Build a model routing layer. Don't let every task hit your most expensive model. Route architecture review to powerful models, route code completion to fast and cheap ones. That alone can bring 40-60% cost reduction at scale.
- Negotiate on volume, not just price. Push for multi-year discounts, free months, custom usage caps, and data processing agreements aligned with your compliance posture.
- Create an internal AI value report every quarter. Present to your stakeholders: productivity delta, cost trends, risk posture. That's how you control AI spend, not just chase it.

Governance
- Put your usage policy in writing. What AI can and can't touch: production secrets, PHI, regulated data.
- Review cost monthly. Actual vs. budget. Anomalies. Vendor commitments.
- Set an acceptance rate threshold. If suggestion acceptance drops below 25%, the tool is noise. Escalate or cut.
- Check your vendor contract before signing: SLA, data retention policy, model version lock, uptime guarantee.

The engineering leaders who build this discipline now will control the AI narrative at their next board review, not explain runaway spend.